energies Article Complementary Analysis for DGA Based on Duval Methods and Furan Compounds Using Artificial Neural Networks 1 1,2, 1, 1,3 Ancut, a-Mihaela Aciu , Claudiu-Ionel Nicola * , Marcel Nicola * and Maria-Cristina Nit, u 1 Research and Development Department, National Institute for Research, Development and Testing in Electrical Engineering-ICMET Craiova, 200746 Craiova, Romania; [email protected] (A.-M.A.); [email protected] (M.-C.N.) 2 Department of Automatic Control and Electronics, University of Craiova, 200585 Craiova, Romania 3 Department Electrical Engineering, Energetic and Aeronautics, University of Craiova, 200585 Craiova, Romania * Correspondence: [email protected] (C.-I.N.); [email protected] (M.N.) Abstract: Power transformers play an important role in electrical systems; being considered the core of electric power transmissions and distribution networks, the owners and users of these assets are increasingly concerned with adopting reliable, automated, and non-invasive techniques to monitor and diagnose their operating conditions. Thus, monitoring the conditions of power transformers has evolved, in the sense that a complete characterization of the conditions of oil–paper insulation can be achieved through dissolved gas analysis (DGA) and furan compounds analysis, since these analyses provide a lot of information about the phenomena that occur in power transformers. The Duval triangles and pentagons methods can be used with a high percentage of correct predictions compared to the known classical methods (key gases, International Electrotechnical Commission (IEC), Rogers, Doernenburg ratios), because, in addition to the six types of basic faults, they also identify four sub-types of thermal faults that provide important additional information for the Citation: Aciu, A.-M.; Nicola, C.-I.; appropriate corrective actions to be applied to the transformers. A new approach is presented based Nicola, M.; Nit,u, M.-C. Complementary Analysis for DGA on the complementarity between the analysis of the gases dissolved in the transformer oil and the Based on Duval Methods and Furan analysis of furan compounds, for the identification of the different faults, especially when there are Compounds Using Artificial Neural multiple faults, by extending the diagnosis of the operating conditions of the power transformers, in Networks. Energies 2021, 14, 588. terms of paper degradation. The implemented software system based on artificial neural networks https://doi.org/10.3390/en14030588 was tested and validated in practice, with good results. Academic Editors: Michel Duval and Keywords: power transformer; insulation; dissolved gas analysis; furan compounds; radial basis Sérgio Cruz function neural network; feed forward neural network Received: 23 November 2020 Accepted: 20 January 2021 Published: 24 January 2021 1. Introduction Publisher’s Note: MDPI stays neutral Power transformers are key pieces of equipment in the electric power transmission with regard to jurisdictional claims in published maps and institutional affil- and distribution systems, and their reliability influences the safety of electric power supply iations. networks. Although they are reliable pieces of equipment, it is difficult to avoid damaging them. In addition to catastrophic damage to the power supply system, the faults in trans- formers can also cause substantial financial losses for both the owner and the consumers served by it. For this reason, it is important to identify, at an early stage, the possible faults in transformers so that, based on an appropriate diagnostic procedure, an efficient and Copyright: © 2021 by the authors. rational decision is taken in advance on the appropriate corrective actions to be applied to Licensee MDPI, Basel, Switzerland. the transformer [1–4]. This article is an open access article distributed under the terms and An important and effective tool for the early-stage fault diagnosis in oil-immersed conditions of the Creative Commons power transformers is dissolved gas analysis (DGA), which can identify the degradation Attribution (CC BY) license (https:// of the solid insulation and oil [5–10]. The diagnosis of the involvement of solid insulation creativecommons.org/licenses/by/ and its possible carbonization, resulting from a method of dissolved gas analysis, can be 4.0/). Energies 2021, 14, 588. https://doi.org/10.3390/en14030588 https://www.mdpi.com/journal/energies Energies 2021, 14, 588 2 of 22 confirmed by careful use of carbon oxides, the CO2/CO ratio, and the analysis of furan compounds [11–15]. Lately, in order to correct and render, more effectively, the methods used for identifying faults in oil-immersed power transformers that rely on the DGA, researchers around the world have tried to apply, to these methods, various techniques based on artificial intelligence, such as the fuzzy logic-based method proposed in [16], as a new solution for determining the fault condition using the combination of gas level, gas rate, and DGA interpretation through the Duval pentagon method. One approach carrying out the classification of power transformer faults, based on the combination of intelligent methods, such as the hypersphere multi-class support vector machine (HMSVM), the hybrid immune a logarithm (HIA), and the kernel extreme learning machine (KELM), is shown in [17]. The optimization of the parameters of the HMSVM- type method, starting from the training stage, is achieved by means of particle swarm optimization (PSO). The fusion of these methods is achieved by the Dempster–Shafer (DS) evidence theory in order to increase the accuracy of the results. Starting from the problem of insufficient and imbalanced datasets, one method that carries out the diagnosis of power transformers and overcomes these issues is the twin support vector machines (TWSVMs) method, proposed in [18]. The parameters of this intelligent classification method are optimized by using the chemical reaction optimization (CRO) type algorithm, and the efficiency and accuracy are increased by means of restricted Boltzmann machines (RBMs) from the moment of data preprocessing. Moreover, to avoid a series of drawbacks regarding the traditional methods of DGA, an approach based on an improved algorithm of the grey wolf type, in order to get better results in case of fault classification by the least square support vector machine (LSSVM), is proposed in [19]. Based on the DGA and the construction of a global health index of power transformers, intelligent algorithms based on the improved differential evolution optimization algorithm, which substantially improve the precision and accuracy in determining the faults, are used in [20,21]. A complementarity for DGA, which is carried out through the partial discharge study by using the correlation analysis and extraction of the main characteristic parameters is shown in [22]. Moreover, a complex image of the fault condition of power transformer is presented in [23] and consists of determining the winding hot spot temperature combined with the DGA. Although the intelligent methods and theories presented above have yielded good results on the accuracy of the power transformer fault diagnosing methods based on the DGA, there are also some shortcomings in the parameter optimization, the selection of the set of characteristics and the data preprocessing methods, which narrow the practical application of the intelligent systems [1,17,18]. The novelty presented in this article consists in the complementarity between the analysis of the gases dissolved in the oil of power transformers and the analysis of furan compounds, for the identification of the different faults, especially when there are multiple faults, by extending the diagnosis of the operating condition of power transformers in terms of paper degradation. This automatic and non-invasive diagnostic is based on the Duval triangles and pentagons analysis methods. The software implementation of the diagnosis is based on using artificial neural networks, such as Radial Basis Function Neural Network (RBFNN) and Feed Forward Neural Network (FFNN), due to their facilities, such as: learning, robustness of the algorithms used, and good results on accuracy and precision. This paper is a continuation of the previous research [24–26] related to the determina- tion of power transformer faults based on the fuzzy logic implementation of the DGA and the analysis of furan compounds. The rest of the paper is organized as follows: Section2 presents the methods for the analysis of dissolved gases and furan compounds. Section3 presents the system developed Energies 2021, 14, 588 3 of 22 to determine the faults in oil-immersed power transformers, based on neural networks. Section4 presents the validation of the proposed system, and the conclusions and future approaches are presented in Section5. 2. Materials and Methods for Faults Detection of the Power Transformers The interpretation of the DGA in oil-filled transformers is one of the most important procedures for the determination of the types of faults. IEC 60599-2015 09 [5] and Institute of Electrical and Electronics Engineers IEEE C57.104-2019 [6] standards provide several DGA interpretation methods achieved on the calculation of certain oil-dissolved flammable gas ratios. Between the methods proposed by these standards and the evolution of the effective faults in transformers, there is a correlation validated by large datasets collected from operating transformers and compared with those collected from out-of-service